The Complete Guide to Building Agents with the Claude Agent SDK

The post explains that the Claude Agent SDK is the engine behind Claude Code, exposed as a library for building agents that can read files, run commands, edit code, manage context, and keep working through tasks. It walks through a code review agent that analyzes a codebase for bugs, security issues, performance problems, and code quality, then returns structured feedback.

Reading notes

#

• The SDK replaces the manual tool loop you would otherwise manage with the raw API.
• Built-in tools include Read, Write, Edit, Bash, Glob, Grep, WebSearch, and WebFetch.
• The guide uses Claude Code CLI as the runtime, TypeScript as the language, and Claude Opus 4.5 as the model.
• Query returns an async generator that streams system, assistant, and result messages.
• The review agent can inspect files, search code, and report issues with file names and line numbers.
• The SDK supports JSON Schema output for structured review results.
• Permission modes include default, acceptEdits, and bypassPermissions.
• canUseTool lets you approve or deny tools with custom logic.
• Subagents can be registered with Task for specialized analysis.
• Sessions can be resumed using a session ID.
• Hooks such as PreToolUse can audit tool calls or block dangerous commands.
• Custom tools can be exposed through MCP servers.
• The guide also covers cost tracking, file checkpointing, skills, hosting, secure deployment, and TypeScript and Python references.